Using Affective Features from Media Content Metadata for Better

Movie Recommendations

John Kalung Leung

, Igor Griva

and William G. Kennedy

Computational and Data Sciences Department, Computational Sciences and Informatics, College of Science,

George Mason University, 4400 University Drive, Fairfax, Virginia 22030, U.S.A.

Department of Mathematical Sciences, MS3F2, Exploratory Hall 4114, George Mason University,

4400 University Drive, Fairfax, Virginia 22030, U.S.A.

Center for Social Complexity, Computational and Data Sciences Department, College of Science,

George Mason University, 4400 University Drive, Fairfax, Virginia 22030, U.S.A.

Keywords: Affective Computing, Affective Aware Top-N Recommendations, Deep Learning, Text-based Emotion

Mining.

Abstract: This paper investigates the causality in the decision making of movie recommendations through the users'

affective profiles. We advocate a method of assigning emotional tags to a movie by the auto-detection of the

affective features in the movie's overview. We apply a text-based Emotion Detection and Recognition model,

which trained by tweets short messages and transfers the learned model to detect movie overviews’ implicit

affective features. We vectorize the affective movie tags to represent the mood embeddings of the movie. We

obtain the user's emotional features by taking the average of all the movies' affective vectors the user has

watched. We apply five-distance metrics to rank the Top-N movie recommendations against the user's

emotion profile. We found Cosine Similarity distance metrics performed better than other distance metrics

measures. We conclude that by replacing the top-N recommendations generated by the Recommender with

the reranked recommendations list made by the Cosine Similarity distance metrics, the user will effectively

get affective aware top-N recommendations while making the Recommender feels like an Emotion Aware

Recommender.

1 INTRODUCTION

Emotion affects human experience and influences our

daily activities on all levels of the decision-making

process. When a user ponders over a list of

recommended items such as songs, books, movies,

products, or services, his affective state of preferences

influences his decision making on which

recommended item he chooses to consume. Emotion

plays a role in our decision-making process in

preference selection (Naqvi et al., 2006). However,

the information retrieval (IF) and Recommender

Systems (RS) field give little attention to include

human emotion as a source of user context (Ho and

Tagmouti, 2006). Our goal in this paper is to make

affective awareness a component in making movie

recommendations for users. The challenge is that no

https://orcid.org/0000-0003-0216-1134

https://orcid.org/0000-0001-9238-1215

film database or movie dataset in the public domain

contains any explicit textual oriented human

emotional tag in the metadata. However, the film

metadata fields such as plot, overview, storyline,

script, watcher reviews, and critics reviews contain

excellent subjective data that describes the general

mood of a movie. We can apply Machine Learning

(ML) techniques to identify and extract affective

features implicitly from the film metadata and

leverage the film's emotional characteristics when

making movie recommendations to users.

No two films are created the same. The moods of

a movie act like an affective fingerprint of the film.

We envision an affective movie feature represents by

a low dimension continuous emotional vector

embedding denotes as the movie's emotional vector

(mvec). Some film databases, such as the MovieLens,

Leung, J., Griva, I. and Kennedy, W.

Using Affective Features from Media Content Metadata for Better Movie Recommendations.

DOI: 10.5220/0010056201610168

In Proceedings of the 12th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2020) - Volume 1: KDIR, pages 161-168

ISBN: 978-989-758-474-9

161

track users' movie-watching history and feedback

(Harper and Konstan, 2016). Using the user's movie-

watching history, we formulate a low dimension

continuous emotion vector embedding denotes as the

user emotional vector (uvec). We obtain a user's uvec

embedding value by taking the average of all the

movies' mvecs the user has watched. Note that uvec

may not be unique if two users watched the same set

of movies. The difference between mvec and uvec is

that mvec of a movie is static, with value unchanged

throughout its lifetime.

In contrast, uvec is dynamic, with its value

changes as the user watched and rated a movie. The

advantage of using the dynamic nature of uvec in the

movie recommendation-making process is that we are

taking the most updated user's affective preference

into consideration of the user's decision-making

process. As the user emotional preference change, the

movie recommender will adjust the recommendation-

making process accordingly. We may be the first

party making use of the novelty in leveraging the

dynamic nature of uvec over mvec to enhance the

movie Recommender recommendation-making

process.

Table 1: Affect values, mvec, of movies "The Godfather

(1972)" derived from balanced and unbalanced moods.

Moods

Balanced

Moods

Dataset

Rank

Unbalanced

Moods

Dataset

Rank

Neutral 0.0840931 6 0.04276474 6

Joy 0.059261046 7 0.16501102 3

Sadness 0.08991193 5 0.076094896 4

Hate 0.23262443 1 0.4305178 1

Anger 0.20177138 2 0.1993026 2

Disgust 0.19720455 3 0.053966276 5

Surprise 0.13513364 4 0.03234269 7

Moreover, we can leverage the range and strength

of a film's moods, i.e., mvec, to analyse a film’s

emotional features. In this study, we track six primary

human affective features in emotion: “joy”,

“sadness”, “hate”, “anger”, “disgust”, and “surprise”.

We added "neutral" as the seventh affective feature

for convenience in affective computation. We

normalized the affective features when we compute

mvec for a film. Thus, all affective features in mvec

will add up to one (1). For example, Internet Movie

Database (IMDb) is a popular online movie

information database that has rated “The Godfather

(1972)” as the top movie of all time (IMDb, 2020).

Our emotion detector classified the movie's dominant

affective class as “hate” and depicted the movie's

mvec in Table 1.

2 RELATED WORK

Detecting primary human emotion expression in text

is a relatively new research area in Natural Language

Processing (NLP). A common approach in

identifying the general thought, feeling, or sense in

writing is to classify the contextual polarity

orientation (positive, neutral, and negative) of

opinionated text through the polarity Sentimental

Analysis (SA) (Wilson et al., 2005), and (Maas et al.,

2011). When applying fine-grained Sentiment

Analysis (Fink et al., 2011), researchers can identify

the intensity level of the polarity as a multi-class

single-label classification problem (e.g., very

positive, optimistic, neutral, negative, and very

negative) (Bhowmick et al., 2009). However, to

determine the mental, emotional state or composure

(i.e., mood) in subjective text, Emotional Analysis

(EA) can better suit to handle the task (Tripathi et al.,

2016). The researcher wants to know the writing

feeling under examination is one of the following

primary human emotions or moods.

The study of basic human emotional expressions

started in the era of Aristotle in around 4th century

BC (Konstan and Konstan, 2006). However, not until

Charles Darwin (1872 – 1998) revisited the

investigation of human emotional expression in the

19th century, which propelled the field to its present

stage of modern psychology research (Ekman, 2006).

Paul Ekman et alia in the 1970s developed a Facial

Action Coding System (FACS) to carry out a series

of research on facial expressions that have identified

the following six primary universal human emotions:

happiness, sadness, disgust, fear, surprise, and anger

(Ekman, 1999). Ekman later added contempt as the

seventh primary human emotion to his list (Ekman et

al., 2013). Robert Plutchik invented the Wheel of

Emotions, advocated eight primary emotions: anger,

anticipation, joy, trust, fear, surprise, sadness, and

disgust. Adding to the primary eight emotions are

secondary and complementary emotions for 32

emotions depicted on the initial Wheel of Emotions

(Plutchik, 2001). More recent research by Glasgow

University in 2014 amended that couple pairs of

emotions such as fear and surprise elicited similar

facial muscles response, so are disgust and anger. The

study broke the raw human emotions down to four

fundamental emotions: happiness, sadness,

fear/surprise, and disgust/anger (Tayib and

Jamaludin, 2016).

Like many researchers have based their work on

Ekman’s six primary human emotions (Canales and

Martínez-Barco, 2014), we also focus our emotion

detection and recognition (EDR) on Ekman’s six

KDIR 2020 - 12th International Conference on Knowledge Discovery and Information Retrieval

162

primary human emotions. We make use of the

WordNet-Affect, a linguistic resource for a lexical

representation of affective knowledge in affective

computing on human interaction such as attention,

emotions, motivation, pleasure, and entertainment

(Valitutti et al., 2004). Emotional expression research

usually aims to detect and recognize emotion types

from human facial expression and vocal intonation

(De Silva et al., 1997). However, our EDR study

focuses on the mood of text expression instead.

Nevertheless, the question remains how much of an

emotion we can convey through writing.

3 METHODOLOGY

In the absence of any publicly available explicit

emotion labeled movie metadata dataset, we build an

affective text aware model in two steps. First, through

readily available tweets data from the Twitter

database, we developed a Tweet Affective Classifier

(TAC). TAC can classify any tweet text into an

affective vector embedding containing seven basic

human emotions in probabilistic values. Next, we

feed the movie text metadata, such as overviews, to

TAC to classify the movie's mvec affective values.

3.1 Data Preparation

One of the challenges in this study is to obtain a large

enough movie metadata set with mood labels. No

such dataset is readily available. We need to build the

required dataset by deriving it from four different

sources:

1. For the movie rating datasets, we obtained the

datasets from the MovieLens datasets stored in

the GroupLens repository (Harper and

Konstan, 2016).

2. We scraped The Movie Database (TMDb)

(TMDb, 2018) for movie overviews and other

metadata.

3. We derived our emotional word sense set as

contextual emotional words synonymous with

WordNet (Miller, 1995).

4. Finally, we scraped the Twitter database for

tweets with keyword tags that matched our

contextual emotion word synonymous (Marres

and Weltevrede, 2013).

MovieLens contains a “links” file containing

cross-reference links between MovieLens’ movie id

and TMDb’s tmdb id. We connect MovieLens and

TMDb datasets through the “links” file.

We build a seven text-based emotion predictor for

movie overviews from the seven-emotion tweet

classifier model. We run the predictor through all the

452,102 overviews scraped from the TMDb database.

We joined the affective aware movie overviews with

four Movielens datasets: ml-latest-small (a.k.a. ml-

latest-small hereafter), ml20m, ml25m and ml-latest

(a.k.a. ml27m hereafter), as depicted in the following

Table 2. Not all movies listed in the MovieLens

datasets has a corresponding movie extracted from

TMDb.

Table 2: MovieLens dataset statistics merged with number

of overviews from TMDb.

Dataset users ratings movies overviews

mlsm 610 100836 9742 9625

ml20m 138493 20000263 27278 26603

ml25m 162641 25000095 62423 60494

ml27m 283228 27753444 58098 56314

3.1.1 Extract Emotion Synonymous from

WordNetAffecteMotionLists

WordNet developed an affective knowledge

linguistic resource known as WordNet-Affect for

lexical representation (Strapparava et al., 2004). The

selection and tagging of a subset of synsets convey

the emotional meaning of a word in WordNet-Affect.

WordNet-Affect emotion lists contain lists of

concepts extracted from WordNet-Affect, synsets

with six emotions of interest: anger, disgust, hate, joy,

sadness, and surprise stored in a compressed file:

“WordNetAffectEmotionLists.tar.gz” (Poria et

al., 2012).

We downloaded the “.gz” file and uncompressed

it into six emotion text files. Each emotion file

contains two columns of information: synsets and the

synonymous. Here, the synonymous set of the synset

corresponds to an emotion class and store in the

corresponding emotion text file. We extract the

synonymous column from each emotion text file. We

removed duplicate synonymous, sorted the cleansed

synonymous, and stored the result in comma-

separated values (CSV) format in the corresponding

emotion synonymous file. Table 3 contains the

statistic of the six emotion synonymous files after

performed the data cleansing task.

Table 3: Synonymous statistics of six emotions.

Mood Count Synonymous List

Joy 400 “admirable”,...,“zestfulness”

Sadness 202 “aggrieve”,...,“wretched”

Hate 147 “affright”,...,“unsure”

Anger 255 “abhor”,...,“wrothful”

Disgust 53 “abhorrent”,...,“yucky”

Surprise 71 “admiration”,...,“wondrously”

Using Affective Features from Media Content Metadata for Better Movie Recommendations

163

3.1.2 Extract Tweets from Twitter Database

There are many types of tweets on Twitter, a popular

social network, and the microblogging platform. This

study only works with the regular tweet, 140

characters, or less short message, which posts on

Twitter. Almost every user’s tweets are extractable

and available to the public. Each tweet is searchable

by keyword. We wrote a simple Python script to

extract tweets through Twitter’s API (Makice, 2009).

We treat each synonymous in an emotion

corresponding file as a keyword of a tweet. By

looping through all the synonymous in Twitter’s

search-by-keyword API, we extract all the tweets

with such keyword and store them in a CSV file. The

alternative is to extract tweets and store them in a

JSON file, as illustrated in (Makice, 2009). For

example, if the emotion synonymous belongs to the

anger concept, we will store the retrieved tweet

in anger raw.csv file. As depicted in Table 3, the

anger emotion corresponding file, the anger syn.txt,

has 255 synonymous. We will store all tweets

retrieved from the corresponding keywords in anger

raw.csv.

After performing text cleansing steps, it yields the

following affective feature records depicted in Table

4. Our affective dataset extracted from Twitter shows

an unbalanced dataset. We decide to balance the

affective dataset by subsampling each affective

attribute dataset size to 15,000. We further split each

affective dataset into a training dataset with 80% of

the samples (12,000) and 20% of the test dataset

(3,000).

Using a brute force method, we scrape the TMDb

database for movie metadata and movie, which

contains the subjective writing movie description to

classify the text’s mood. Our effort yields 452,102

records after cleansing the raw data we scraped from

TMDb.

Table 4: Moods records gathered from Twitter.

Mood Class Size

Neutral 19180

Joy 138019

Sadness 60381

Hate 38651

Anger 17830

Disgust 19887

Surprise 15002

No. of unbalance 7 mood classes 308878

No. of each balanced mood 15000

No. of each balanced mood train 12000

No. of each balanced mood test 3000

3.2 Emotion Modeling

Intrigue by the recent publication in Natural

Language Processing (NLP) for text classification

described by (Sosa, 2017), we develop a text-based

EDR model by concatenating Long Short-Term

Memory (LSTM) architecture and Conv-1D of

Convolutional Neural Network (CNN) architecture.

We follow a similar method used in (Liu, 2020) to

build our EDR model. We define our model

architecture consists of two halves. The first half is

RNN LSTM-CNN Conv-1D architecture, as

described in (Sosa, 2017) that text input process by an

LSTM architecture before following up data

processing by a CNN Conv-1D architecture. In

contrast, the second half of the model is to reverse the

processing order of architecture, CNN Conv-1D-

RNN LSTM. Input first process by a CNN Conv-1D

architecture before feeding it to an RNN LSTM

architecture. The two halves of the architecture then

combine to feed data into a max-pooling layer of a

CNN for a pooling operation to select the dominant

feature in the filter’s regional feature map. Next, data

passes into a CNN flattening layer to convert data into

a one-dimensional array before feed data to a fully

connected CNN layer. The dense layer’s output will

feed to a set of nodes equal to the number of classes

the architecture aims to classify. Each of the output

nodes holds the output distribution value of its class.

In the final act, a softmax activation function examine

and activate the appropriate class node accordingly.

We use bi-directional RNN LSTM and CNN

Conv1D architectures to build our model. In the first

half of the model, the RNN LSTM-CNN Conv-1D

phase, we use two bi-directional LSTM for the RNN

LSTM architecture and seven Conv1D CNN

architecture. In the second half of the model, the

CNN-LSTM phase, we apply seven pairs of Conv1D

of the CNN architecture and two bi-directional LSTM

for the LSTM architecture. Follow the idea illustrated

in (Kim, 2014); we prepared two identical input

layers of embedding matrix constructed from a pre-

trained GloVe embedding matric similar to

(Pennington et al., 2014). We build the two input

layers of the embedding matrix with one of the input

embedding layers set to “trainable,” while the other is

not, i.e., “frozen.” During the processing of the first

half of the model, the RNN LSTM-CNN Conv-1D

phase, the “trainable” input layer occupies one of the

bi-directional LSTM architectures. In contrast, the

“frozen” input layer fills the other.

Similarly, when processing the second half of the

model, the CNN Conv-1D – RNN LSTM phase, a

“trainable” input layer, and a “frozen” input layer will

KDIR 2020 - 12th International Conference on Knowledge Discovery and Information Retrieval

164

Table 5: Emotion Detection Recognition on balanced

moods tweets.

Precision Recall F1-score Support

Neutral 0.47 0.77 0.59 2992

Joy 0.63 0.53 0.58 3030

Sadness 0.64 0.44 0.52 3034

Hate 0.64 0.51 0.57 2933

Anger 0.62 0.68 0.65 2984

Disgust 0.44 0.45 0.44 2987

Surprise 0.55 0.51 0.53 3040

Accuracy 0.56 21000

Macro avg 0.57 0.56 0.55 21000

Weighted avg 0.57 0.56 0.55 21000

occupy each pair of the Conv1D units. We obtain

55.6% accuracy in classifying the emotion class of the

tweets’ balanced dataset, as depicted in Table 5, and

the confusion matrix in figure 1 depicts the

performance of the seven-emotion classifier. The

classification result is acceptable to serve our purpose

since our goal is not to build the best emotion text

classifier, but a usable one to classify the emotion

class of movie overviews.

Figure 1: Confusion matrix of 7 balanced emotions dataset.

4 IMPLEMENTATIONS

4.1 UVEC and MVEC

We selected all users in the MovieLens ml-latest-

small dataset as test users. We split the number of

movies watched by each user into the user training set

and validation set. Table 6 illustrates of the

computation results of uvecs from a 20-80% split of

train and validation sets. Table 8 depicts an example

of user id 400 uvec statistics. We sorted user id 400

by the timestamp attribute in the ml-latest-small

ratings data file to simulate the movie watching order.

We take the 20% split train set to compute the uvec

for the test user. We used the test user’s last watched

movie, “Lucky Number Slevin (2006),” as input to

the SVD-CFRS Recommender to generate a top-20

recommendations list. We applied the five-distance

metrics against the top-20 recommendation list to

obtain five corresponding re-ranked top-20

recommendation lists, as depicted in Table 8. We then

compared the number of movies matched in each

recommendation list against the movies in the

validation set of the to-be-watched list to report as the

hit percentage.

Table 6: MLSM dataset test user uvec, movie watched

count, 20% train and 80% validate.

User ID 1 2 … 610

Wcount 2698 1864 … 2108

x%train …

y%validate …

UVEC

Neutral

Joy

Sadness

Hate

Anger

Disgust

Surprise

0.16635188

0.09730581

0.1180924

0.1641951

0.11517799

0.17250315

0.16637367

0.17283309

0.09685813

0.11604573

0.16120733

0.11227607

0.17098578

0.16979389

…

0.16885831

0.09974659

0.1187206

0.16087716

0.11261272

0.17191968

0.16726495

4.2 Comparative Platform

We developed a movie Recommender System to

evaluate the performance of user emotion profile,

uvec, and movie emotion profile, mvec. We envision

uvec and mvec play a role in the tail end process of

making movie top-N Recommendations. Any movie

Recommenders can refit to support uvec and mvec

processes. We developed an SVD based

Collaborative Filtering movie Recommender System

(SVD-CFRS). We added functions to support uvec

and mvec operations as enhancement of making

movie recommendations.

The five-distance metrics we employed in the

comparative platform were Euclidean distance,

Manhattan distance, Minkowski distance, Cosine

similarity, and Pearson correlation with their formula

illustrated in equation 1 through 6.





x,y



















1

(1)





x,y















1

(2)

Using Affective Features from Media Content Metadata for Better Movie Recommendations

165





x,y

















1





(3)





x,y













〈

,

〉



(4)

CosSim







∑



∑



∑



x,









(5)













,



(6)

Where (x, y) are vectors x = (x

, … , x

) and y =

, …, y

5 EVALUATION

5.1 Findings

We derived the following methods to evaluate the

performance of the recommendations list generated

by the SVD-CF Recommender and the reranking of

the recommendations list by the five-distance metrics

algorithms. We join the ml-latest-small dataset's

ratings and movie data files with the emotion TMDb

dataset to obtain the movie dataset with emotion

labels. Table 7 illustrates the pseudo-code of

grouping users and sorted their activity by timestamp

to simulate their movie watching order. The pseudo-

code also shows splitting each user's activity into a

test and validation set and compute uvec. It also

shows in the pseudo-code the top-20, top-10, and top-

5 of recommendations lists and the reranking of the

recommendations list by the five-distance metrics.

Lastly, the pseudo-code computes the hit rate in

percent for top-20, top-10, and top-5

recommendations.

Table 7: Pseudocode of getting users uvecs, topN, and hit

rates.

1 Group user by userId

2 For each user do

3 Sort user row by timestamp

4 Split user into test and validation set: 20/80

5 Get uvec for test user

6 Get top20/10/5 by test user’s last watched film

7 Rerank top20 by 5 distance metrics, get top10/5

8 Get hit% by validation set for top20/10/5

9 Get hit% by validation set for 5-dist top20/10/5

Table 8 shows an example of test user id 400 of last

watched movie statistics and the movie's mvec.

Table 8: Test user id 400 last watched movie and mvec.

uwerId 400 tmdbId 186

movieId 44665 rating 4.0

title Lucky

Number

Slevin

(2006)

genres Crime,

Drama,

Mystery

timestamp 1498870148 m_neutral 0.204775

m_happy 0.075458 m_sad 0.127180

m_hate 0.171851 m_anger 0.112534

m_disgust 0.166984 m_surprise 0.141217

Table 9 depicts a 20-80 split sampling result of the

test user id 400. The column "Mid" shows the movie

ids of top-20, top-10, and top-5 recommendations list

generated by the SVD-CF Recommender and the

recommendations lists' reranking by the five-distance

metrics. Also, the table shows the hit rates of top-20,

top-10, and top-5 for each category.

Table 9: Sampling on 20-80 split of 43 watched movies of

ml-latest-small user id 400 into 8 watched movies for uvec

computation and 35 to-be-watched movies for validation on

the top20 and top5 recommendations list generated by

SVD-CFRS and five distance.

400 Mid Euc Mht Mki Cos Pear

1 296 58559 58559 58559 2959 2959

2 50 7153 7153 7153 2329 58559

3 858 356 5952 356 527 2028

4 2959 5952 356 5952 2858 2329

5 593 593 1089 593 1213 527

6 4993 1089 593 1089 858 858

7 58559 1221 1221 608 79132 1089

8 7153 608 50 1221 2028 2858

9 608 4993 4993 4993 47 1213

10 1221 2028 2028 2028 50 47

11 527 296 608 296 296 79132

12 79132 50 296 47 4993 4993

13 1213 47 47 79132 608 1221

14 5952 858 858 50 1089 7153

15 2858 79132 79132 858 1221 296

16 356 1213 2858 1213 593 608

17 47 2858 527 2858 5952 593

18 2329 527 2329 527 356 5952

19 1089 2329 1213 2329 58559 50

20 2028 2959 2959 2959 7153 356

T20% 70 70 70 70 70 70

T5% 100 80 60 80 40 40

T10% 100 70 70 70 60 50

5.2 Find the Winner of TopN Ranked

by Distance Metrics

Table 10 shows the average categorical hit rate of all

test users. The first column, “Mid” shows the average

hit rate obtained through various split sizes of top-20,

top-10, and top-5 for the non-affective aware

KDIR 2020 - 12th International Conference on Knowledge Discovery and Information Retrieval

166

recommendations lists. The other five columns show

the affective aware hit rate results. To make the

regular non-affective aware recommender become an

effective aware Recommender, choose the winner

from the five-distance metric. In our case, the Cosine

Similarity hit rate as shown in the “Cos” column

yields the best result across the top-20, top-10, and

top-5 when comparing among the others.

Table 10: Sampling on various split sizes for each test user

to obtain the respective recommendations list and the

reranking recommendations lists by five-distance metrics

and hit rates on each user’s category for reporting the

average results.

Top20

Split

Mid

Hit%

Euc

Hit%

Mht

Hit%

Mki

Hit%

Cos

Hit%

Pear

Hit%

10-90 69.34 69.34 69.34 69.34 69.34 69.34

20-80 64.98 64.98 64.98 64.98 64.98 64.98

30-70 59.86 59.86 59.86 59.86 59.86 59.86

40-60 54.24 54.24 54.24 54.24 54.24 54.24

50-50 48.42 48.42 48.42 48.42 48.42 48.42

60-40 41.53 41.53 41.53 41.53 41.53 41.53

70-30 34.37 34.37 34.37 34.37 34.37 34.37

80-20 25.75 25.75 25.75 25.75 25.75 25.75

90-10 16.42 16.42 16.42 16.42 16.42 16.42

Top5

Split

Mid

Hit%

Euc

Hit%

Mht

Hit%

Mki

Hit%

Cos

Hit%

Pear

Hit%

10-90 82.26 69.41 69.41 69.67 69.80 69.80

20-80 78.20 64.30 64.13 64.00 65.87 65.57

30-70 72.46 59.57 59.67 59.51 60.23 60.23

40-60 67.11 54.10 55.02 54.33 54.33 54.03

50-50 60.98 49.84 49.51 49.38 48.10 47.87

60-40 53.34 42.13 42.03 41.87 40.82 40.79

70-30 44.03 34.46 34.69 34.59 34.69 34.13

80-20 34.56 25.48 25.54 25.28 25.41 26.07

90-10 21.80 17.11 16.67 17.18 15.41 16.00

Top10

Split

Mid

Hit%

Euc

Hit%

Mht

Hit%

Mki

Hit%

Cos

Hit%

Pear

Hit%

10-90 76.89 68.82 68.97 68.93 69.89 69.54

20-80 72.49 64.26 64.23 64.05 65.66 65.08

30-70 67.31 59.33 59.41 59.44 60.41 60.00

40-60 61.59 53.90 54.08 54.03 54.54 54.43

50-50 55.13 48.57 48.74 48.59 48.18 48.30

60-40 47.77 41.92 42.02 41.93 41.13 41.39

70-30 39.66 34.64 34.66 34.52 34.13 34.00

80-20 29.93 25.54 25.85 25.60 25.90 25.75

90-10 19.28 16.54 16.66 16.54 16.25 16.08

6 FUTURE WORK

We plan to elaborate our work in mvec and uvec from

making the Recommender become emotion aware of

the recommendations making process, a bottom-up

approach, to build an Emotion Aware Recommender

from the top down. We also plan to use affective

features in users’ emotion profiles to enhance Group

Recommender in group formation, group dynamics,

and group decision-making.

7 CONCLUSIONS

We illustrate a strategy to transform a non-affective

aware Recommender to become affective aware. We

started by developing an Emotion Detection and

Recognition (EDR) model to classify seven

emotional features in tweets through emotion tags

stored in the Twitter database. We then transferred the

EDR model's learning from classifying the affective

features of tweets to classify a movie's emotion

through the movie overview. We generated emotional

features, mvec, for each collected movie from TMDb

and joined it with the ratings dataset found in the

MovieLens repository. We use the emotion labeled

ratings dataset to make the top-N recommendations

list through an SVD-CF Recommender while adding

functions to support uvec and mvec for affective

computing and analysis. We reranked all test

users'users' top-N recommendations lists through

five-distance metrics. We systematically evaluated

the recommendations lists' performance and the

reranked recommendation lists and found Cosine

Similarity distance metrics performed the best. We

conclude that the Cosine Similarity algorithm is the

most suitable distance metrics to use in making a non-

affective aware SVD-CF Recommender affective

aware.

REFERENCES

Bhowmick, P.K., Basu, A., Mitra, P., Prasad, A., 2009.

Multi-label text classification approach for sentence

level news emotion analysis, in: International

Conference on Pattern Recognition and Machine

Intelligence. Springer, pp. 261–266.

Canales, L., Martínez-Barco, P., 2014. Emotion detection

from text: A survey, in: Proceedings of the Workshop

on Natural Language Processing in the 5th Information

Systems Research Working Days (JISIC). pp. 37–43.

De Silva, L.C., Miyasato, T., Nakatsu, R., 1997. Facial

emotion recognition using multi-modal information, in:

Proceedings of ICICS, 1997 International Conference

on Information, Communications and Signal

Processing. Theme: Trends in Information Systems

Engineering and Wireless Multimedia

Communications (Cat. IEEE, pp. 397–401.

Ekman, P., 2006. Darwin and facial expression: A century

of research in review. Ishk.

Ekman, P., 1999. Basic emotions. Handbook of cognition

and emotion 98, 16.

Ekman, P., Friesen, W.V., Ellsworth, P., 2013. Emotion in

the human face: Guidelines for research and an

integration of findings. Elsevier.

Fink, C.R., Chou, D.S., Kopecky, J.J., Llorens, A.J., 2011.

Coarse- and Fine-Grained Sentiment Analysis of Social

Using Affective Features from Media Content Metadata for Better Movie Recommendations

167

Media Text. Johns hopkins apl technical digest 30, 22–

30.

Harper, F.M., Konstan, J.A., 2016. The movielens datasets:

History and context. ACM Transactions on Interactive

Intelligent Systems (TiiS) 5, 19.

http://dx.doi.org/10.1145/2827872

Ho, I.L., Ai Thanh, Menezes, Tagmouti, Y., 2006. E-mrs:

Emotion-based movie recommender system, in:

Proceedings of IADIS E-Commerce Conference. USA:

University of Washington Both-Ell. pp. 1–8.

IMDb, 2020. Top 100 Greatest Movies of All Time (The

Ultimate List) [WWW Document]. URL

http://www.imdb.com/list/ls055592025/ (accessed

5.17.20).

Kim, Y., 2014. Convolutional neural networks for sentence

classification. arXiv preprint arXiv:1408.5882.

Konstan, D., Konstan, D., 2006. The emotions of the

ancient Greeks: Studies in Aristotle and classical

literature. University of Toronto Press.

Liu, T., 2020. Multi-class Emotion Classification for Short

Texts [WWW Document]. URL

https://tlkh.github.io/text-emotion-classification/

(accessed 5.18.20).

Maas, A.L., Daly, R.E., Pham, P.T., Huang, D., Ng, A.Y.,

Potts, C., 2011. Learning Word Vectors for Sentiment

Analysis, in: Proceedings of the 49th Annual Meeting

of the Association for Computational Linguistics:

Human Language Technologies. Association for

Computational Linguistics, Portland, Oregon, USA, pp.

142–150.

Makice, K., 2009. Twitter API: Up and running: Learn how

to build applications with the Twitter API. O’Reilly

Media, Inc.

Marres, N., Weltevrede, E., 2013. Scraping the social?

Issues in live social research. Journal of cultural

economy 6, 313–335.

Miller, G.A., 1995. WordNet: a lexical database for

English. Communications of the ACM 38, 39–41.

Naqvi, N., Shiv, B., Bechara, A., 2006. The role of emotion

in decision making: A cognitive neuroscience

perspective. Current directions in psychological science

15, 260–264.

Pennington, J., Socher, R., Manning, C.D., 2014. Glove:

Global vectors for word representation, in: Proceedings

of the 2014 Conference on Empirical Methods in

Natural Language Processing (EMNLP). pp. 1532–

1543.

Plutchik, R., 2001. The nature of emotions: Human

emotions have deep evolutionary roots, a fact that may

explain their complexity and provide tools for clinical

practice. American scientist 89, 344–350.

Poria, S., Gelbukh, A., Cambria, E., Yang, P., Hussain, A.,

Durrani, T., 2012. Merging SenticNet and WordNet-

Affect emotion lists for sentiment analysis, in: 2012

IEEE 11th International Conference on Signal

Processing. IEEE, pp. 1251–1255.

Sosa, P.M., 2017. Twitter Sentiment Analysis using

Combined LSTM-CNN Models. Eprint Arxiv.

Strapparava, C., Valitutti, A., others, 2004. Wordnet affect:

an affective extension of wordnet., in: Lrec. Citeseer, p.

40.

Tayib, S., Jamaludin, Z., 2016. An Algorithm to Define

Emotions Based on Facial Gestures as Automated Input

in Survey Instrument. Advanced Science Letters 22,

2889–2893.

TMDb, 2018. TMDb About [WWW Document]. URL

https://www.themoviedb.org/about?language=en

(accessed 5.11.18).

Tripathi, V., Joshi, A., Bhattacharyya, P., 2016. Emotion

analysis from text: A survey. Center for Indian

Language Technology Surveys.

Valitutti, A., Strapparava, C., Stock, O., 2004. Developing

affective lexical resources. PsychNology Journal 2, 61–

83.

Wilson, T., Wiebe, J., Hoffmann, P., 2005. Recognizing

contextual polarity in phrase-level sentiment analysis,

in: Proceedings of Human Language Technology

Conference and Conference on Empirical Methods in

Natural Language Processing. pp. 347–354.

KDIR 2020 - 12th International Conference on Knowledge Discovery and Information Retrieval

168